Considerations in using recurrent neural networks to probe neural dynamics

Area of Science:

• Computational neuroscience
• Machine learning in biology
• Neural network modeling

Background:

• Recurrent neural networks (RNNs) are increasingly used to model animal behavior and neural activity.
• RNNs offer an in silico model for understanding neural computations in cortical areas.

Purpose of the Study:

• To investigate considerations for using RNNs to model motor behavior in delayed reach tasks.
• To explore how network properties affect the capture of single-neuron versus population-level neural motifs.
• To analyze the dynamical mechanisms underlying RNNs and their impact on robustness and generalization.

Main Methods:

• Trained RNNs to reproduce animal behavior and neural activity statistics.
• Varied network nonlinear activation and rate regularization to assess motif capture.
• Analyzed network dynamics through low-dimensional projections.
• Tested RNN robustness to input noise and generalization to untrained tasks.

Main Results:

• RNNs reproducing single-neuron firing rates may not capture population-level motifs.
• Different training methods can lead to distinct dynamical mechanisms in RNNs, even when reproducing similar neurophysiological features.
• RNNs with specific dynamics demonstrated increased robustness to input noise.
• Trained RNNs generalized to tasks with altered targets or locations, indicating sufficient learned dynamics.

Conclusions:

• Emphasizes the importance of considering both single-neuron and population-level dynamics when using RNNs for neural modeling.
• Highlights that RNNs can exhibit different underlying mechanisms for computation, necessitating careful analysis.
• Demonstrates the potential of RNNs to generalize, providing insights into neural computation and robustness.